Merge remote-tracking branch 'origin/perf-report-ins-number' into bert-opt2

Dockerfile

@@ -86,7 +86,7 @@ RUN git clone --single-branch --branch ${ONNXRUNTIME_BRANCH} --recursive ${ONNXR
 
 ADD tools/build_and_test_onnxrt.sh /onnxruntime/build_and_test_onnxrt.sh
 
-RUN PATH=/opt/cmake/bin:$PATH cget -p /usr/local install ROCmSoftwarePlatform/llvm-project-mlir@02078ce236ad90e3aec04c0c770ef5bfc99e49c2
+RUN cget -p /usr/local install ROCmSoftwarePlatform/llvm-project-mlir@26a4b3cfc0a1a15181490f24ae461608fef1b04e -DBUILD_MIXR_TARGET=On
 
 ENV MIOPEN_FIND_DB_PATH=/tmp/miopen/find-db
 ENV MIOPEN_USER_DB_PATH=/tmp/miopen/user-db

src/include/migraphx/matcher.hpp

@@ -349,25 +349,27 @@ match::matcher_result find_match(module& modl, M&& m)
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TRACE_MATCHES)
 
 /// Find matches for an instruction in the module
-template <class... Ms>
-void find_matches(module& mod, instruction_ref ins, Ms&&... ms)
+template <class Mod, class... Ms>
+void find_matches(Mod& mod, instruction_ref ins, Ms&&... ms)
 {
 #if !defined(__GNUC__) || defined(__clang__) || __GNUC__ > 5
     const
 #endif
-        bool trace = enabled(MIGRAPHX_TRACE_MATCHES{});
+        int trace = value_of(MIGRAPHX_TRACE_MATCHES{});
     bool match    = false;
     each_args(
         [&](auto&& m) {
             if(match)
                 return;
-            auto r = match_instruction(mod, ins, m.matcher());
-            if(r.result == mod.end())
+            if(trace > 1)
+                std::cout << "Match: " << get_type_name(m) << std::endl;
+            auto r = match_instruction(get_module(mod), ins, m.matcher());
+            if(r.result == get_module(mod).end())
                 return;
-            if(trace)
+            if(trace > 0)
             {
                 std::cout << "Matched by " << get_type_name(m) << std::endl;
-                mod.debug_print(ins);
+                get_module(mod).debug_print(ins);
             }
             m.apply(mod, r);
             match = true;
@@ -376,10 +378,10 @@ void find_matches(module& mod, instruction_ref ins, Ms&&... ms)
 }
 
 /// Find matches in a module
-template <class... Ms>
-void find_matches(module& mod, Ms&&... ms)
+template <class Mod, class... Ms>
+void find_matches(Mod& mod, Ms&&... ms)
 {
-    for(auto ins : iterator_for(mod))
+    for(auto ins : iterator_for(get_module(mod)))
     {
         find_matches(mod, ins, ms...);
     }

src/include/migraphx/module.hpp

@@ -124,7 +124,7 @@ struct module
                      std::unordered_map<instruction_ref, instruction_ref> map_ins = {});
 
     std::vector<instruction_ref>
-    add_instructions(module_ref m,
+    add_instructions(const_module_ref m,
                      std::unordered_map<instruction_ref, instruction_ref> map_ins = {});
 
     std::vector<instruction_ref>
@@ -139,7 +139,7 @@ struct module
 
     std::vector<instruction_ref>
     insert_instructions(instruction_ref ins,
-                        module_ref m,
+                        const_module_ref m,
                         std::unordered_map<instruction_ref, instruction_ref> map_ins = {});
 
     std::vector<instruction_ref>
@@ -227,6 +227,8 @@ struct module
     std::unique_ptr<module_impl> impl;
 };
 
+inline module& get_module(module& m) { return m; }
+
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx
 

src/include/migraphx/module_ref.hpp

@@ -33,6 +33,7 @@ inline namespace MIGRAPHX_INLINE_NS {
 
 struct module;
 using module_ref       = module*;
+using const_module_ref = const module*;
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/include/migraphx/ranges.hpp

@@ -198,6 +198,12 @@ void transform(Range&& r, Iterator it, F f)
     std::transform(r.begin(), r.end(), it, f);
 }
 
+template <class Range1, class Range2, class Iterator, class F>
+void transform(Range1&& r1, Range2&& r2, Iterator it, F f)
+{
+    std::transform(r1.begin(), r1.end(), r2.begin(), it, f);
+}
+
 template <class Range>
 auto reverse(Range& r)
 {

src/include/migraphx/shape.hpp

@@ -191,6 +191,10 @@ struct shape
 
         std::size_t size(std::size_t n = 1) const { return sizeof(type) * n; }
 
+        auto is_integral() const { return std::is_integral<type>{}; }
+        auto is_signed() const { return std::is_signed<type>{}; }
+        auto is_unsigned() const { return std::is_unsigned<type>{}; }
+
         template <class U>
         type* from(U* buffer, std::size_t n = 0) const
         {

src/module.cpp

@@ -399,7 +399,8 @@ module::add_instructions(const std::vector<instruction_ref>& instructions,
 }
 
 std::vector<instruction_ref>
-module::add_instructions(module_ref m, std::unordered_map<instruction_ref, instruction_ref> map_ins)
+module::add_instructions(const_module_ref m,
+                         std::unordered_map<instruction_ref, instruction_ref> map_ins)
 {
     return this->insert_instructions(this->end(), m, std::move(map_ins));
 }
@@ -420,8 +421,10 @@ module::insert_instructions(instruction_ref ins,
     return insert_generic_instructions(*this, ins, instructions, std::move(map_ins));
 }
 
-std::vector<instruction_ref> module::insert_instructions(
-    instruction_ref ins, module_ref m, std::unordered_map<instruction_ref, instruction_ref> map_ins)
+std::vector<instruction_ref>
+module::insert_instructions(instruction_ref ins,
+                            const_module_ref m,
+                            std::unordered_map<instruction_ref, instruction_ref> map_ins)
 {
     return insert_generic_instructions(*this, ins, iterator_for(*m), std::move(map_ins));
 }

src/program.cpp

@@ -681,11 +681,13 @@ void program::perf_report(std::ostream& os,
     double overhead_percent       = overhead_time * 100.0 / total_time;
     double total_instruction_time = 0.0;
     std::unordered_map<std::string, double> op_times;
+    std::unordered_map<std::string, std::size_t> op_n;
     for(auto&& p : ins_vec)
     {
         double avg = common_average(p.second);
         op_times[perf_group(p.first->get_operator())] += avg;
         total_instruction_time += avg;
+        op_n[perf_group(p.first->get_operator())]++;
     }
     double calculate_overhead_time    = total_time - total_instruction_time;
     double calculate_overhead_percent = calculate_overhead_time * 100.0 / total_time;
@@ -706,18 +708,19 @@ void program::perf_report(std::ostream& os,
 
     os << std::endl;
     os << "Summary:" << std::endl;
-    std::vector<std::pair<double, std::string>> op_times_sorted;
-    std::transform(op_times.begin(),
-                   op_times.end(),
-                   std::back_inserter(op_times_sorted),
-                   [](auto p) { return std::make_pair(p.second, p.first); });
+    std::vector<std::tuple<double, std::size_t, std::string>> op_times_sorted;
+    std::transform(
+        op_times.begin(), op_times.end(), std::back_inserter(op_times_sorted), [&](auto p) {
+            auto&& name = p.first;
+            return std::make_tuple(p.second, op_n.at(name), name);
+        });
     std::sort(op_times_sorted.begin(), op_times_sorted.end(), std::greater<>{});
-    for(auto&& p : op_times_sorted)
+    for(auto&& [avg, nn, name] : op_times_sorted)
     {
-        auto&& name    = p.second;
-        double avg     = p.first;
         double percent = std::ceil(100.0 * avg / total_instruction_time);
-        os << name << ": " << avg << "ms, " << percent << "%" << std::endl;
+        double per_ins = avg / nn;
+        os << name << ": " << avg << "ms / " << nn << " = " << per_ins << "ms, " << percent << "%"
+           << std::endl;
     }
 
     os << std::endl;

src/targets/gpu/CMakeLists.txt

@@ -164,6 +164,7 @@ add_library(migraphx_gpu
     deconvolution.cpp
     device_name.cpp
     elu.cpp
+    fuse_mlir.cpp
     fuse_ops.cpp
     gather.cpp
     gemm_impl.cpp
@@ -176,7 +177,7 @@ add_library(migraphx_gpu
     loop.cpp
     lrn.cpp
     leaky_relu.cpp
-    mlir_conv.cpp
+    mlir.cpp
     multinomial.cpp
     nonzero.cpp
     pack_args.cpp
@@ -320,16 +321,26 @@ message(STATUS "extractkernel: ${MIGRAPHX_EXTRACT_KERNEL}")
 
 set(MIGRAPHX_ENABLE_MLIR OFF CACHE BOOL "")
 if(MIGRAPHX_ENABLE_MLIR)
-    find_library(LIBMLIRMIOPEN MLIRMIOpenThin REQUIRED)
+    find_library(MLIRAPI_LIBRARY MLIRMIOpen 
+        PATH_SUFFIXES
+        # Workaournd broken mlir install
+        lib/ lib/lib)
     # REQUIRED is not supported before cmake 3.18
-    if(NOT LIBMLIRMIOPEN)
-        message(FATAL_ERROR "libMLIRMIOpenThin not found")
+    if(NOT MLIRAPI_LIBRARY)
+        message(FATAL_ERROR "libMLIRMIOpen not found")
     else()
-        message(STATUS "Build with libMLIRMIOpenThin: " ${LIBMLIRMIOPEN})
+        message(STATUS "Build with libMLIRMIOpen: " ${MLIRAPI_LIBRARY})
     endif()
 
-    target_compile_definitions(migraphx_gpu PRIVATE "-DMIGRAPHX_MLIR_MIOPEN_SUPPORT")
-    target_link_libraries(migraphx_gpu PUBLIC ${LIBMLIRMIOPEN})
+    find_path(MLIRAPI_HEADERS NAMES mlir-c/Dialect/MIGraphX.h)
+    # Workaround MLIR broken installation
+    find_path(MLIRAPI_HEADERS2 NAMES mlir-c/Registration.h
+        PATH_SUFFIXES 
+        include/external/include external/include)
+
+    target_compile_definitions(migraphx_gpu PRIVATE "-DMIGRAPHX_MLIR")
+    target_include_directories(migraphx_gpu SYSTEM PRIVATE ${MLIRAPI_HEADERS} ${MLIRAPI_HEADERS2})
+    target_link_libraries(migraphx_gpu PUBLIC ${MLIRAPI_LIBRARY})
 endif()
 
 set(MIGRAPHX_USE_HIPRTC OFF CACHE BOOL "")

src/targets/gpu/code_object_op.cpp

@@ -52,7 +52,7 @@ code_object_op::compute(context& ctx, const shape&, const std::vector<argument>&
     std::transform(
         args.begin(), args.end(), kargs.begin(), [](const argument& a) { return a.data(); });
     k.launch(ctx.get_stream().get(), global, local, std::move(kargs));
-    return args.back();
+    return args[get_output_arg(args.size())];
 }
 void code_object_op::finalize(context&, const shape&, const std::vector<shape>&)
 {

src/targets/gpu/fuse_mlir.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/gpu/fuse_mlir.hpp>
+#include <migraphx/gpu/mlir.hpp>
+#include <migraphx/matcher.hpp>
+#include <migraphx/pass_manager.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/register_op.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct module;
+
+namespace gpu {
+
+#ifdef MIGRAPHX_MLIR
+struct mlir_conv
+{
+    operation op = make_op("convolution");
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.op, "op"));
+    }
+
+    std::string name() const { return "gpu::mlir_conv"; }
+    shape compute_shape(std::vector<shape> inputs, const std::vector<module_ref>& mods) const
+    {
+        check_shapes{inputs, *this}.standard();
+        if(mods.size() != 1)
+            MIGRAPHX_THROW("should have one submodule.");
+        if(inputs.size() < 2)
+            MIGRAPHX_THROW("should have at least two inputs.");
+        auto n = inputs.size();
+        return op.compute_shape({inputs[n - 2], inputs[n - 1]});
+    }
+};
+MIGRAPHX_REGISTER_OP(mlir_conv);
+
+namespace {
+struct find_conv_pointwise
+{
+    // Find a convolution followed by a pointwise operation.
+    auto matcher() const
+    {
+        auto convolution =
+            match::skip(match::name("contiguous"))(match::name("convolution").bind("convolution"));
+        return match::name("pointwise")(match::any_of[match::inputs()](convolution.bind("x")));
+    }
+
+    void apply(module_pass_manager& mpm, const match::matcher_result& r) const
+    {
+        auto ins      = r.result;
+        auto conv_ins = r.instructions["convolution"];
+        auto x_ins    = r.instructions["x"]; // input after contiguous
+        auto* pm      = ins->module_inputs().front();
+        auto names    = pm->get_parameter_names();
+        // Whitelist pointwise operators
+        if(std::any_of(pm->begin(), pm->end(), [](const auto& i) {
+               return not contains({"@literal", "@param", "@return", "convolution", "add", "relu"},
+                                   i.name());
+           }))
+            return;
+        // Only fuse with fp32 for now
+        if(std::any_of(ins->inputs().begin(), ins->inputs().end(), [&](auto i) {
+               return i->get_shape().type() != shape::type_t::float_type;
+           }))
+            return;
+        std::sort(names.begin(), names.end());
+        module_ref mm = mpm.create_module("mlir_" + pm->name());
+        mm->set_bypass();
+        std::unordered_map<instruction_ref, instruction_ref> param_map;
+        auto x    = mm->add_parameter("x" + std::to_string(names.size()),
+                                   conv_ins->inputs().at(0)->get_shape());
+        auto w    = mm->add_parameter("x" + std::to_string(names.size() + 1),
+                                   conv_ins->inputs().at(1)->get_shape());
+        auto conv = mm->add_instruction(conv_ins->get_operator(), {x, w});
+        std::transform(names.begin(),
+                       names.end(),
+                       ins->inputs().begin(),
+                       std::inserter(param_map, param_map.end()),
+                       [&](auto name, auto input) {
+                           if(input == x_ins)
+                               return std::make_pair(pm->get_parameter(name), conv);
+                           return std::make_pair(pm->get_parameter(name),
+                                                 mm->add_parameter(name, input->get_shape()));
+                       });
+        mm->add_return(mm->insert_instructions(mm->end(), pm, param_map));
+
+        std::vector<instruction_ref> inputs;
+        std::copy_if(ins->inputs().begin(),
+                     ins->inputs().end(),
+                     std::back_inserter(inputs),
+                     [&](auto input) { return input != conv_ins; });
+        inputs.insert(inputs.end(), conv_ins->inputs().begin(), conv_ins->inputs().end());
+        mpm.get_module().replace_instruction(
+            ins, mlir_conv{conv_ins->get_operator()}, inputs, {mm});
+    }
+};
+} // namespace
+
+#endif
+
+void fuse_mlir::apply(module_pass_manager& mpm) const
+{
+#ifdef MIGRAPHX_MLIR
+    match::find_matches(mpm, find_conv_pointwise{});
+#else
+    (void)mpm;
+#endif
+}
+
+} // namespace gpu
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/fuse_ops.cpp

@@ -342,6 +342,7 @@ void move_standard_front(std::vector<instruction_ref>& args)
 
 auto gpu_name(const std::string& s) { return match::name("gpu::" + s); }
 
+namespace {
 struct find_layernorm
 {
     auto matcher() const { return match::layernorm(&gpu_name); }
@@ -843,15 +844,6 @@ inline auto precompile_name(Strings... names) // NOLINT
     });
 }
 
-template <class... Ms>
-auto conv_bias_pointwise(Ms... ms)
-{
-    return precompile_name("pointwise")(
-        match::either_arg(0, 1)(bias_shape(match::used_once()).bind("bias"),
-                                fusable_conv(match::used_once()).bind("conv")),
-        ms...);
-}
-
 struct find_conv_bias
 {
     context* ctx = nullptr;
@@ -1087,6 +1079,7 @@ struct find_commutative_broadcast
         m.replace_instruction(ins, ins->get_operator(), args);
     }
 };
+} // namespace
 
 struct find_contiguous
 {

src/targets/gpu/include/migraphx/gpu/code_object_op.hpp

@@ -38,12 +38,13 @@ struct context;
 
 struct code_object_op
 {
-    value::binary code_object;
-    std::string symbol_name;
-    std::size_t global;
-    std::size_t local;
-    std::vector<shape> expected_inputs;
-    shape output;
+    value::binary code_object{};
+    std::string symbol_name = "";
+    std::size_t global      = 0;
+    std::size_t local       = 0;
+    std::vector<shape> expected_inputs{};
+    shape output{};
+    std::int64_t output_arg = -1;
     kernel k{};
 
     template <class Self, class F>
@@ -66,9 +67,13 @@ struct code_object_op
     argument
     compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
     void finalize(context&, const shape&, const std::vector<shape>&);
+    std::int64_t get_output_arg(std::size_t n) const
+    {
+        return output_arg < 0 ? n + output_arg : output_arg;
+    }
     std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
     {
-        return shapes.size() - 1;
+        return get_output_arg(shapes.size());
     }
 
     friend std::ostream& operator<<(std::ostream& os, const code_object_op& op)

src/targets/gpu/include/migraphx/gpu/mlir_conv.hpp → src/targets/gpu/include/migraphx/gpu/fuse_mlir.hpp

@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  * THE SOFTWARE.
  */
-#ifndef MIGRAPHX_GUARD_RTGLIB_MIOPEN_MLIR_CONV_HPP
-#define MIGRAPHX_GUARD_RTGLIB_MIOPEN_MLIR_CONV_HPP
+#ifndef MIGRAPHX_GUARD_GPU_FUSE_MLIR_HPP
+#define MIGRAPHX_GUARD_GPU_FUSE_MLIR_HPP
 
 #include <migraphx/config.hpp>
 #include <migraphx/gpu/context.hpp>
@@ -30,18 +30,19 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-struct module;
+struct module_pass_manager;
 
 namespace gpu {
-struct mlir_conv
+
+struct fuse_mlir
 {
-    context* ctx;
-    std::string name() const { return "mlir::convolution"; }
-    void apply(module& m) const;
+    context* ctx = nullptr;
+    std::string name() const { return "gpu::fuse_mlir"; }
+    void apply(module_pass_manager& mpm) const;
 };
 
 } // namespace gpu
+
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx
-
-#endif
+#endif // MIGRAPHX_GUARD_GPU_FUSE_MLIR_HPP

src/targets/gpu/include/migraphx/gpu/mlir.hpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#ifndef MIGRAPHX_GUARD_RTGLIB_GPU_MLIR_HPP
+#define MIGRAPHX_GUARD_RTGLIB_GPU_MLIR_HPP
+
+#include <string>
+#include <vector>
+#include <migraphx/config.hpp>
+#include <migraphx/gpu/code_object_op.hpp>
+#include <migraphx/instruction_ref.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+struct module;
+namespace gpu {
+
+std::string dump_mlir(const module& m);
+code_object_op compile_mlir(const context& ctx, const module& m);
+
+instruction_ref insert_mlir(module& m,
+                            instruction_ref ins,
+                            code_object_op co,
+                            const std::vector<instruction_ref>& inputs);
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/targets/gpu/jit/mlir.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/gpu/compiler.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/gpu/context.hpp>
+
+#include <migraphx/gpu/mlir.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct mlir_compiler : compiler<mlir_compiler>
+{
+    std::vector<std::string> names() const { return {"gpu::mlir_conv"}; }
+
+    operation compile_op(context&, const std::vector<shape>&, const value&) const { return {}; }
+
+    compiler_replace compile(context& ctx, instruction_ref ins, const operation&) const
+    {
+        auto* smod = ins->module_inputs().front();
+        assert(smod->get_parameter_names().size() == ins->inputs().size() - 1);
+        return insert(compile_mlir(ctx, *smod));
+    }
+
+    compiler_replace insert(code_object_op co) const
+    {
+        return [co = std::move(co)](module& m, instruction_ref ins) {
+            auto mlir = insert_mlir(m, ins, co, ins->inputs());
+            m.replace_instruction(ins, mlir);
+        };
+    }
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/mlir.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/gpu/mlir.hpp>
+
+#ifdef MIGRAPHX_MLIR
+#include <mlir-c/IR.h>
+#include <mlir-c/BuiltinAttributes.h>
+#include <mlir-c/BuiltinTypes.h>
+#include <mlir-c/Diagnostics.h>
+#include <mlir-c/Dialect/MIGraphX.h>
+#include <mlir-c/IntegerSet.h>
+#include <mlir-c/Pass.h>
+#include <mlir-c/Registration.h>
+#endif
+
+#include <migraphx/env.hpp>
+#include <migraphx/manage_ptr.hpp>
+#include <migraphx/module.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/config.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/gpu/code_object_op.hpp>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/device_name.hpp>
+#include <migraphx/iterator_for.hpp>
+#include <deque>
+#include <variant>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TRACE_MLIR);
+
+#ifdef MIGRAPHX_MLIR
+template <class T, class F, F f> // NOLINT
+struct mlir_handle
+{
+    struct ptr
+    {
+        ptr() = default;
+        ptr(std::nullptr_t) {}
+        ptr(T x) : obj(x) {}
+
+        std::intptr_t get_value() const
+        {
+            static_assert(sizeof(T) == sizeof(std::intptr_t), "MLIR Handle different size");
+            return reinterpret_cast<const std::intptr_t&>(obj);
+        }
+
+        T get() const { return obj; }
+
+        friend bool operator==(ptr x, ptr y) { return x.get_value() == y.get_value(); }
+
+        friend bool operator!=(ptr x, ptr y) { return !(x == y); }
+        T obj{};
+    };
+
+    struct deleter
+    {
+        using pointer = ptr;
+
+        void operator()(pointer x) const
+        {
+            if(x != nullptr)
+            {
+                (void)f(x.obj);
+            }
+        }
+    };
+
+    mlir_handle() : handle(nullptr) {}
+
+    mlir_handle(T p) : handle(ptr{p}) {}
+
+    T get() const { return handle.get().get(); }
+
+    T release() { return handle.release().get(); }
+
+    private:
+    std::unique_ptr<ptr, deleter> handle;
+};
+
+#define MIGRAPHX_MANAGE_MLIR_HANDLE(T, F) migraphx::gpu::mlir_handle<T, decltype(&F), &F> // NOLINT
+
+using mlir_context           = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirContext, mlirContextDestroy);
+using mlir_module            = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirModule, mlirModuleDestroy);
+using mlir_operation         = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirOperation, mlirOperationDestroy);
+using mlir_op_printing_flags = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirOpPrintingFlags,
+                                                           mlirOpPrintingFlagsDestroy);
+using mlir_region            = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirRegion, mlirRegionDestroy);
+using mlir_block             = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirBlock, mlirBlockDestroy);
+using mlir_pass_manager      = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirPassManager, mlirPassManagerDestroy);
+
+std::string_view to_string_view(MlirStringRef s) { return {s.data, s.length}; }
+
+MlirStringRef make_mlir_string_ref(const std::string_view& s)
+{
+    return mlirStringRefCreate(s.data(), s.size());
+}
+
+template <class F, class T, class Printer>
+void mlir_print(F f, T x, Printer printer)
+{
+    f(
+        x,
+        +[](MlirStringRef s, void* data) {
+            (*reinterpret_cast<Printer*>(data))(to_string_view(s));
+        },
+        &printer);
+}
+
+template <class F, class T>
+void mlir_print(F f, T x, std::ostream& os)
+{
+    mlir_print(f, x, [&](auto s) { os << s; });
+}
+
+template <class F, class T>
+std::string mlir_print(F f, T x)
+{
+    std::stringstream ss;
+    mlir_print(f, x, [&](auto s) { ss << s; });
+    return ss.str();
+}
+
+struct mlir_program
+{
+    mlir_program()
+        : ctx(mlirContextCreate()),
+          location(mlirLocationUnknownGet(ctx.get())),
+          mmodule(mlirModuleCreateEmpty(location))
+    {
+        MlirDialectHandle mixr_handle = mlirGetDialectHandle__migraphx__();
+        mlirDialectHandleRegisterDialect(mixr_handle, ctx.get());
+        mlirRegisterAllDialects(ctx.get());
+        mlirContextSetAllowUnregisteredDialects(ctx.get(), true /*allow*/);
+    }
+
+    MlirType make_type(shape::type_t t) const
+    {
+        MlirType result;
+        shape::visit(t, [&](auto as) {
+            if(as.type_enum() == shape::float_type)
+                result = mlirF32TypeGet(ctx.get());
+            else if(as.type_enum() == shape::half_type)
+                result = mlirF16TypeGet(ctx.get());
+            else if(as.type_enum() == shape::double_type)
+                result = mlirF64TypeGet(ctx.get());
+            else if(as.is_integral())
+            {
+                if(as.is_signed())
+                    result = mlirIntegerTypeSignedGet(ctx.get(), as.size() * 8);
+                else
+                    result = mlirIntegerTypeGet(ctx.get(), as.size() * 8);
+            }
+            else
+                MIGRAPHX_THROW("Unsupported type: " + std::to_string(as.type_enum()));
+        });
+        return result;
+    }
+
+    MlirType make_tensor(const shape& s) const
+    {
+        assert(s.standard());
+        std::vector<int64_t> lens(s.lens().begin(), s.lens().end());
+        return mlirRankedTensorTypeGet(
+            lens.size(), lens.data(), make_type(s.type()), mlirAttributeGetNull());
+    }
+
+    template <class Range>
+    std::vector<MlirType> make_tensors(const Range& r)
+    {
+        std::vector<MlirType> result;
+        std::transform(r.begin(), r.end(), std::back_inserter(result), [&](const auto& s) {
+            return make_tensor(s);
+        });
+        return result;
+    }
+
+    MlirType make_function_type(const std::vector<shape>& inputs, const std::vector<shape>& outputs)
+    {
+        auto in  = make_tensors(inputs);
+        auto out = make_tensors(outputs);
+        return mlirFunctionTypeGet(ctx.get(), in.size(), in.data(), out.size(), out.data());
+    }
+
+    MlirIdentifier id(const std::string_view& s) const
+    {
+        return mlirIdentifierGet(ctx.get(), make_mlir_string_ref(s));
+    }
+
+    MlirAttribute attribute(std::int64_t i) const
+    {
+        if(i < 0)
+            MIGRAPHX_THROW("MLIR cant handle negative values since they are ambiguous");
+        return mlirIntegerAttrGet(mlirIntegerTypeGet(ctx.get(), 64), i);
+    }
+    MlirAttribute attribute(std::uint64_t i) const
+    {
+        if(i > (std::numeric_limits<std::uint64_t>::max() / 2))
+            MIGRAPHX_THROW("MLIR cant handle large integer values since they are ambiguous");
+        return mlirIntegerAttrGet(mlirIntegerTypeGet(ctx.get(), 64), i);
+    }
+    MlirAttribute attribute(unsigned char i) const { return attribute(std::uint64_t(i)); }
+    MlirAttribute attribute(bool b) const { return mlirBoolAttrGet(ctx.get(), b ? 1 : 0); }
+    MlirAttribute attribute(double d) const
+    {
+        return mlirFloatAttrDoubleGet(ctx.get(), mlirF64TypeGet(ctx.get()), d);
+    }
+    MlirAttribute attribute(const std::string& s) const
+    {
+        return mlirStringAttrGet(ctx.get(), make_mlir_string_ref(s));
+    }
+    MlirAttribute attribute(std::nullptr_t) const { return {}; }
+    template <class T>
+    MlirAttribute attribute(const std::vector<T>& v) const
+    {
+        std::vector<MlirAttribute> attributes;
+        attributes.reserve(v.size());
+        std::transform(v.begin(), v.end(), std::back_inserter(attributes), [&](auto&& x) {
+            return attribute(x);
+        });
+        return mlirArrayAttrGet(ctx.get(), attributes.size(), attributes.data());
+    }
+    MlirAttribute attribute(const value& v) const
+    {
+        MlirAttribute attr;
+        v.visit_value([&](auto&& x) { attr = attribute(x); });
+        return attr;
+    }
+    MlirAttribute attribute(const std::vector<value>& v) const
+    {
+        if(v.empty())
+        {
+            return mlirArrayAttrGet(ctx.get(), 0, nullptr);
+        }
+        if(not v.front().get_key().empty())
+        {
+            std::vector<MlirNamedAttribute> attributes = name_attributes(v);
+            return mlirDictionaryAttrGet(ctx.get(), attributes.size(), attributes.data());
+        }
+        else
+        {
+            std::vector<MlirAttribute> attributes;
+            attributes.reserve(v.size());
+            std::transform(v.begin(), v.end(), std::back_inserter(attributes), [&](auto&& x) {
+                return attribute(x);
+            });
+            return mlirArrayAttrGet(ctx.get(), attributes.size(), attributes.data());
+        }
+    }
+
+    MlirAttribute attribute(MlirType t) const { return mlirTypeAttrGet(t); }
+
+    MlirAttribute attribute(MlirAttribute a) const { return a; }
+
+    template <class T>
+    MlirNamedAttribute name_attribute(const std::string_view& key, const T& x) const
+    {
+        MlirNamedAttribute attr;
+        attr.name      = id(key);
+        attr.attribute = attribute(x);
+        return attr;
+    }
+
+    using attribute_t       = std::variant<std::nullptr_t,
+                                     std::uint64_t,
+                                     unsigned char,
+                                     bool,
+                                     double,
+                                     std::string,
+                                     value,
+                                     std::vector<value>,
+                                     MlirType>;
+    using named_attribute_t = std::pair<std::string_view, attribute_t>;
+
+    MlirNamedAttribute name_attribute(const named_attribute_t& na) const
+    {
+        return name_attribute(na.first,
+                              std::visit([&](const auto& x) { return attribute(x); }, na.second));
+    }
+
+    std::vector<MlirNamedAttribute>
+    name_attributes(const std::vector<named_attribute_t>& named_attrs) const
+    {
+        std::vector<MlirNamedAttribute> attributes;
+        attributes.reserve(named_attrs.size());
+        std::transform(named_attrs.begin(),
+                       named_attrs.end(),
+                       std::back_inserter(attributes),
+                       [&](const named_attribute_t& a) { return name_attribute(a); });
+        return attributes;
+    }
+
+    std::vector<MlirNamedAttribute> name_attributes(const value& v) const
+    {
+        std::vector<MlirNamedAttribute> attributes;
+        attributes.reserve(v.size());
+        std::transform(v.begin(), v.end(), std::back_inserter(attributes), [&](const value& x) {
+            return name_attribute(x.get_key(), x.without_key());
+        });
+        return attributes;
+    }
+
+    struct mlir_operation_state
+    {
+        mlir_operation_state(mlir_program& p, const std::string_view& name)
+            : prog(&p), op_state(mlirOperationStateGet(make_mlir_string_ref(name), p.location))
+        {
+        }
+
+        mlir_operation_state& add_attributes(const std::vector<named_attribute_t>& named_attrs)
+        {
+            auto attributes = prog->name_attributes(named_attrs);
+            mlirOperationStateAddAttributes(&op_state, attributes.size(), attributes.data());
+            return *this;
+        }
+
+        mlir_operation_state& add_attribute_value(const value& v)
+        {
+            auto attributes = prog->name_attributes(v);
+            mlirOperationStateAddAttributes(&op_state, attributes.size(), attributes.data());
+            return *this;
+        }
+
+        mlir_operation_state& add_regions(std::vector<mlir_region> rs)
+        {
+            regions = std::move(rs);
+            return *this;
+        }
+
+        mlir_operation_state& add_region(mlir_region r)
+        {
+            regions.emplace_back(std::move(r));
+            return *this;
+        }
+
+        mlir_operation_state& add_results(const std::vector<shape>& outputs)
+        {
+            auto x = prog->make_tensors(outputs);
+            mlirOperationStateAddResults(&op_state, x.size(), x.data());
+            return *this;
+        }
+
+        mlir_operation_state& add_operands(const std::vector<MlirValue>& inputs)
+        {
+            mlirOperationStateAddOperands(&op_state, inputs.size(), inputs.data());
+            return *this;
+        }
+
+        mlir_operation create_operation()
+        {
+            std::vector<MlirRegion> mregions(regions.size());
+            std::transform(regions.begin(), regions.end(), mregions.begin(), [](const auto& r) {
+                return r.get();
+            });
+            mlirOperationStateAddOwnedRegions(&op_state, mregions.size(), mregions.data());
+            mlir_operation op(mlirOperationCreate(&op_state));
+            // Release memory since mlir_operation owns it
+            for(auto& r : regions)
+                r.release();
+            regions.clear();
+            return op;
+        }
+
+        mlir_program* prog;
+        MlirOperationState op_state;
+        std::vector<mlir_region> regions = {};
+    };
+
+    mlir_operation_state create_operation_state(const std::string_view& name)
+    {
+        return {*this, name};
+    }
+
+    std::vector<MlirValue> insert(MlirBlock body, mlir_operation_state ops)
+    {
+        std::vector<MlirValue> result;
+        mlir_operation op = ops.create_operation();
+        auto weak_op      = op.get();
+        mlirBlockAppendOwnedOperation(body, op.release());
+
+        auto n = mlirOperationGetNumResults(weak_op);
+        result.reserve(n);
+        transform(range(n), std::back_inserter(result), [&](auto i) {
+            return mlirOperationGetResult(weak_op, i);
+        });
+        return result;
+    }
+
+    MlirBlock
+    insert(MlirBlock body, const module& m, std::unordered_map<instruction_ref, MlirValue>& ins_map)
+    {
+        auto names = m.get_parameter_names();
+        std::sort(names.begin(), names.end());
+        std::vector<shape> inputs;
+        std::transform(names.begin(),
+                       names.end(),
+                       std::back_inserter(inputs),
+                       [&](const std::string& name) { return m.get_parameter_shape(name); });
+        std::vector<shape> outputs = m.get_output_shapes();
+
+        std::vector<MlirLocation> arg_locs(inputs.size(), location);
+        auto body_inputs   = make_tensors(inputs);
+        mlir_region region = mlirRegionCreate();
+        mlir_block fbody = mlirBlockCreate(body_inputs.size(), body_inputs.data(), arg_locs.data());
+        MlirBlock result = fbody.get();
+        mlirRegionAppendOwnedBlock(region.get(), fbody.release());
+
+        auto ops = create_operation_state("func.func");
+        ops.add_attributes({{"function_type", make_function_type(inputs, outputs)},
+                            {"sym_name", std::string("main")},
+                            {"kernel", std::string("mixr")}});
+        ops.add_region(std::move(region));
+        insert(body, std::move(ops));
+
+        for(auto i : range(names.size()))
+            ins_map[m.get_parameter(names[i])] = mlirBlockGetArgument(result, i);
+        return result;
+    }
+
+    static std::string get_name(instruction_ref ins)
+    {
+        if(ins->name() == "@return")
+            return "func.return";
+        return "migraphx." + ins->name();
+    }
+
+    static value get_operator_value(const operation& op)
+    {
+        auto v = op.to_value();
+        if(op.name() == "convolution")
+        {
+            // Adjust symetrical padding
+            if(v.at("padding").size() == v.at("stride").size())
+            {
+                auto padding = v.at("padding");
+                std::copy(padding.begin(), padding.end(), std::back_inserter(v.at("padding")));
+            }
+        }
+        return v;
+    }
+
+    static shape get_shape(instruction_ref ins)
+    {
+        if(ins->name() == "@return")
+        {
+            assert(ins->inputs().size() == 1);
+            return ins->inputs().front()->get_shape();
+        }
+        return ins->get_shape();
+    }
+
+    void parse(const module& m)
+    {
+        auto mbody = mlirModuleGetBody(mmodule.get());
+        std::unordered_map<instruction_ref, MlirValue> ins_map;
+        auto fbody = insert(mbody, m, ins_map);
+        for(auto ins : iterator_for(m))
+        {
+            if(ins->name() == "@param")
+                continue;
+            auto name = get_name(ins);
+            auto ops  = create_operation_state(name);
+            ops.add_attribute_value(get_operator_value(ins->get_operator()));
+            if(ins->name() != "@return")
+                ops.add_results({get_shape(ins)});
+
+            std::vector<MlirValue> inputs;
+            transform(
+                ins->inputs(), std::back_inserter(inputs), [&](auto i) { return ins_map.at(i); });
+            ops.add_operands(inputs);
+
+            auto outputs = insert(fbody, std::move(ops));
+            if(ins->name() != "@return")
+            {
+                assert(outputs.size() == 1);
+                ins_map[ins] = outputs.front();
+            }
+        }
+    }
+
+    code_object_op compile() MIGRAPHX_TIDY_CONST
+    {
+        mlir_pass_manager pm{mlirPassManagerCreate(ctx.get())};
+        // 1st pipeline to call
+        mlirMIGraphXAddHighLevelPipeline(pm.get());
+        // 2nd pipeline to call
+        std::string tname = get_device_name();
+        // HACK: Since MLIR can't handle the full target name
+        auto hacked_tname = tname.substr(0, tname.find(':'));
+        if(tname.size() != hacked_tname.size())
+            std::cout
+                << "*************** WARNING: MLIR may not compile the correct target features for: "
+                << tname << std::endl;
+        mlirMIGraphXAddBackendPipeline(pm.get(), hacked_tname.c_str(), "amdgcn-amd-amdhsa", "");
+        mlirPassManagerRun(pm.get(), mmodule.get());
+
+        code_object_op op{};
+        op.symbol_name                = "main";
+        op.code_object                = get_binary();
+        std::tie(op.global, op.local) = get_launch_params();
+        return op;
+    }
+
+    std::pair<std::size_t, std::size_t> get_launch_params() const
+    {
+        uint32_t attrs[2];
+        // returns block and grid sizes
+        mlirGetKernelAttrs(mmodule.get(), attrs);
+        std::size_t local  = attrs[0];
+        std::size_t global = local * attrs[1];
+        return {global, local};
+    }
+
+    value::binary get_binary() const
+    {
+        int size = 0;
+        mlirGetBinary(mmodule.get(), &size, nullptr);
+        value::binary result(size);
+        if(mlirGetBinary(mmodule.get(), &size, reinterpret_cast<char*>(result.data())))
+            return result;
+        MIGRAPHX_THROW("Failed to compile mlir program");
+    }
+
+    mlir_context ctx;
+    MlirLocation location;
+    mlir_module mmodule;
+    std::deque<std::string> strings{};
+};
+
+std::string dump_mlir(const module& m)
+{
+    mlir_program mp;
+    mp.parse(m);
+    auto mod_op = mlirModuleGetOperation(mp.mmodule.get());
+    return mlir_print(&mlirOperationPrint, mod_op);
+}
+
+code_object_op compile_mlir(const context&, const module& m)
+{
+    const bool trace = enabled(MIGRAPHX_TRACE_MLIR{});
+    if(trace)
+        std::cout << m << std::endl;
+    mlir_program mp;
+    mp.parse(m);
+    auto mod_op = mlirModuleGetOperation(mp.mmodule.get());
+    if(trace)
+        std::cout << mlir_print(&mlirOperationPrint, mod_op) << std::endl;
+    auto co   = mp.compile();
+    co.output = m.get_output_shapes().front();
+    return co;
+}
+
+instruction_ref insert_mlir(module& m,
+                            instruction_ref ins,
+                            code_object_op co,
+                            const std::vector<instruction_ref>& inputs)
+{
+    std::vector<instruction_ref> refs;
+    refs.reserve(inputs.size() * 15);
+
+    std::unordered_map<uint64_t, instruction_ref> literal_map{};
+    auto get_literal = [&](uint64_t value) {
+        auto fi = literal_map.find(value);
+        if(fi != literal_map.end())
+            return fi->second;
+        auto lit = m.add_literal(value);
+        literal_map.emplace(value, lit);
+        return lit;
+    };
+
+    std::size_t last = 0;
+    for(auto input : inputs)
+    {
+        const size_t offset = 0;
+        auto s              = input->get_shape();
+        last                = refs.size();
+        refs.push_back(input);
+        refs.push_back(input);
+        refs.push_back(get_literal(offset)); // offset
+
+        // dim sizes
+        std::transform(s.lens().begin(),
+                       s.lens().end(),
+                       std::back_inserter(refs),
+                       [&](const auto& lval) { return get_literal(lval); });
+        // refs.push_back(get_literal(1)); // G
+
+        // dim strides
+        std::transform(s.strides().begin(),
+                       s.strides().end(),
+                       std::back_inserter(refs),
+                       [&](const auto& lval) { return get_literal(lval); });
+        // refs.push_back(get_literal(1)); // G
+    }
+    co.expected_inputs = to_shapes(refs);
+    co.output_arg      = last;
+    return m.insert_instruction(ins, co, refs);
+}
+
+#else
+
+std::string dump_mlir(const module&) { return {}; }
+
+code_object_op compile_mlir(const context&, const module&) { return {}; }
+
+template <class T>
+void use(T&)
+{
+}
+
+instruction_ref
+// cppcheck-suppress funcArgNamesDifferent
+insert_mlir(module& m, instruction_ref, code_object_op co, const std::vector<instruction_ref>&)
+{
+    use(co);
+    return m.end();
+}
+
+#endif
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/mlir_conv.cpp

-/*
- * The MIT License (MIT)
- *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in
- * all copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
- * THE SOFTWARE.
- */
-#include <migraphx/gpu/mlir_conv.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/make_op.hpp>
-
-#include <migraphx/op/convolution.hpp>
-
-#include <migraphx/gpu/context.hpp>
-#include <migraphx/gpu/convolution.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/program.hpp>
-
-#include <migraphx/ranges.hpp>
-#include <migraphx/make_op.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/program.hpp>
-#include <migraphx/gpu/kernel.hpp>
-#include <migraphx/gpu/target.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/gpu/compile_hip.hpp>
-
-#include <utility>
-#include <functional>
-#include <algorithm>
-
-#ifdef MIGRAPHX_MLIR_MIOPEN_SUPPORT
-#include <Miir.h>
-#endif // MIGRAPHX_MLIR_MIOPEN_SUPPORT
-
-#include <cstdio>
-
-namespace migraphx {
-inline namespace MIGRAPHX_INLINE_NS {
-namespace gpu {
-
-struct mlir_apply
-{
-    module* mod           = nullptr;
-    const mlir_conv* pass = nullptr;
-
-    const char* mlir_kernel_name = "migraphx_conv2d";
-
-    std::unordered_map<uint64_t, instruction_ref> literal_map{};
-
-    struct execution_spec
-    {
-        migraphx::value::binary binary;
-        size_t global_size;
-        size_t local_size;
-        execution_spec(migraphx::value::binary&& binary_m, size_t global_s, size_t local_s)
-            : binary(std::move(binary_m)), global_size(global_s), local_size(local_s)
-        {
-        }
-    };
-
-    std::unordered_map<std::string, std::shared_ptr<execution_spec>> binary_map{};
-
-    context& get_context() const
-    {
-        assert(pass != nullptr);
-        assert(pass->ctx != nullptr);
-        return *pass->ctx;
-    }
-
-    void init() const
-    {
-        assert(mod != nullptr);
-        assert(pass != nullptr);
-    }
-
-    std::shared_ptr<execution_spec> make_mlir_binary(instruction_ref op_r)
-    {
-        std::shared_ptr<execution_spec> result;
-
-#ifdef MIGRAPHX_MLIR_MIOPEN_SUPPORT
-        auto conv  = any_cast<op::convolution>(op_r->get_operator());
-        auto inp_t = op_r->inputs().at(0)->get_shape();
-        auto flt_t = op_r->inputs().at(1)->get_shape();
-        auto out_t = op_r->get_shape();
-
-        auto get_type_str = [](const shape& s) -> const char* {
-            switch(s.type())
-            {
-            case shape::float_type: return "f32";
-            case shape::half_type: return "f16";
-            case shape::bool_type:
-            case shape::double_type:
-            case shape::uint8_type:
-            case shape::int8_type:
-            case shape::uint16_type:
-            case shape::int16_type:
-            case shape::int32_type:
-            case shape::int64_type:
-            case shape::uint32_type:
-            case shape::uint64_type:
-            case shape::tuple_type: break;
-            }
-            return nullptr;
-        };
-
-        const auto* inp_t_s = get_type_str(inp_t);
-        const auto* flt_t_s = get_type_str(flt_t);
-        const auto* out_t_s = get_type_str(out_t);
-
-        if(out_t_s == nullptr || inp_t_s == nullptr || flt_t_s == nullptr)
-            return result;
-
-        std::string mlir_options = "--kernel_name " + std::string(mlir_kernel_name);
-
-        // platform spec
-        auto& device = get_context().get_current_device();
-        char dev_name[64];
-        sprintf(dev_name, "gfx%lu%02lu", device.get_device_major(), device.get_device_minor());
-        mlir_options += " --arch " + std::string(dev_name) + " --num_cu " +
-                        std::to_string(device.get_cu_count()); // ???
-
-        // Conv spec
-        mlir_options +=
-            " --operation "
-            "conv2d"
-            " --batchsize " +
-            std::to_string(conv.group) + " --groupsize " + std::to_string(1) + " --padding_h " +
-            std::to_string(conv.padding[0]) + " --padding_w " + std::to_string(conv.padding[1]) +
-            " --conv_stride_h " + std::to_string(conv.stride[0]) + " --conv_stride_w " +
-            std::to_string(conv.stride[1]) + " --dilation_h " + std::to_string(conv.dilation[0]) +
-            " --dilation_w " + std::to_string(conv.dilation[1]);
-
-        // Input spec
-        mlir_options += " --in_layout "
-                        "NCHWG"
-                        " --in_type " +
-                        std::string(inp_t_s) + " --in_channels " + std::to_string(inp_t.lens()[1]) +
-                        " --in_h " + std::to_string(inp_t.lens()[2]) + " --in_w " +
-                        std::to_string(inp_t.lens()[3]);
-
-        // Filter spec
-        mlir_options += " --fil_layout "
-                        "NCHWG"
-                        " --fil_type " +
-                        std::string(flt_t_s) + " --fil_h " + std::to_string(flt_t.lens()[2]) +
-                        " --fil_w " + std::to_string(flt_t.lens()[3]);
-
-        // Output spec
-        mlir_options += " --out_layout "
-                        "NCHWG"
-                        " --out_type " +
-                        std::string(out_t_s) + " --out_channels " +
-                        std::to_string(out_t.lens()[1]) + " --out_h " +
-                        std::to_string(out_t.lens()[2]) + " --out_w " +
-                        std::to_string(out_t.lens()[3]);
-
-        auto bin_i = binary_map.find(mlir_options);
-        if(bin_i == binary_map.end())
-        {
-            size_t bin_size = 0;
-
-            using mlir_handle = MIGRAPHX_MANAGE_PTR(MiirHandle, miirDestroyHandle);
-            auto handle       = mlir_handle(miirCreateHandle(mlir_options.c_str()));
-
-            if(miirLowerBin(handle.get()) == MIIR_SUCCESS &&
-               miirBufferGet(handle.get(), nullptr, &bin_size) == MIIR_SUCCESS)
-            {
-                migraphx::value::binary bin(bin_size);
-                if(miirBufferGet(handle.get(), reinterpret_cast<char*>(bin.data()), &bin_size) ==
-                   MIIR_SUCCESS)
-                {
-                    size_t global_size;
-                    size_t block_size;
-                    if(miirGetExecutionDims(handle.get(), &global_size, &block_size) ==
-                       MIIR_SUCCESS)
-                    {
-                        result = std::make_shared<execution_spec>(
-                            std::move(bin), global_size, block_size);
-                    }
-                }
-            }
-
-            binary_map[mlir_options] = result;
-        }
-        else
-        {
-            result = bin_i->second;
-        }
-#else  // MIGRAPHX_MLIR_MIOPEN_SUPPORT
-        (void)op_r;
-#endif // MIGRAPHX_MLIR_MIOPEN_SUPPORT
-        return result;
-    }
-
-    instruction_ref get_literal(uint64_t value)
-    {
-        auto fi = literal_map.find(value);
-        if(fi != literal_map.end())
-            return fi->second;
-        auto lit = mod->add_literal(value);
-        literal_map.emplace(value, lit);
-        return lit;
-    }
-
-    operation make_code_object_op(instruction_ref op_r, const std::shared_ptr<execution_spec>& spec)
-    {
-        // each pointer is expanded out to a MemRefDescriptor
-        auto inp_t = op_r->inputs().at(0)->get_shape();
-        auto flt_t = op_r->inputs().at(1)->get_shape();
-        auto out_t = op_r->get_shape();
-
-        auto i64 = shape(shape::uint64_type);
-
-        std::vector<shape> expected_inputs = {
-            flt_t, flt_t, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64,  i64,   inp_t,
-            inp_t, i64,   i64, i64, i64, i64, i64, i64, i64, i64, i64, i64,  out_t, out_t,
-            i64,   i64,   i64, i64, i64, i64, i64, i64, i64, i64, i64, out_t};
-
-        return migraphx::make_op("gpu::code_object",
-                                 {
-                                     {"code_object", spec->binary},
-                                     {"symbol_name", mlir_kernel_name},
-                                     {"global", spec->global_size},
-                                     {"local", spec->local_size},
-                                     {"expected_inputs", migraphx::to_value(expected_inputs)},
-                                     {"output", migraphx::to_value(out_t)},
-                                 });
-    }
-
-    void add_memref_descriptor(std::vector<instruction_ref>& refs, instruction_ref inst)
-    {
-        const size_t offset = 0;
-        auto inst_t         = inst->get_shape();
-        refs.push_back(inst);
-        refs.push_back(inst);
-        refs.push_back(get_literal(offset)); // offset
-
-        // dim sizes
-        std::transform(inst_t.lens().begin(),
-                       inst_t.lens().end(),
-                       std::back_inserter(refs),
-                       [&](const auto& lval) { return get_literal(lval); });
-        refs.push_back(get_literal(1)); // G
-
-        // dim strides
-        std::transform(inst_t.strides().begin(),
-                       inst_t.strides().end(),
-                       std::back_inserter(refs),
-                       [&](const auto& lval) { return get_literal(lval); });
-        refs.push_back(get_literal(1)); // G
-    }
-
-    instruction_ref insert_allocation(instruction_ref ins, const shape& s) const
-    {
-        return mod->insert_instruction(ins, hip_allocate{s});
-    }
-
-    void replace_conv_op(instruction_ref ins)
-    {
-        auto conv_bin = make_mlir_binary(ins);
-        if(conv_bin)
-        {
-            auto conv = make_code_object_op(ins, conv_bin);
-
-            auto inp = ins->inputs().at(0);
-            auto flt = ins->inputs().at(1);
-            auto out = insert_allocation(ins, ins->get_shape());
-
-            std::vector<instruction_ref> refs;
-            refs.reserve(3 * 13 + 1);
-            add_memref_descriptor(refs, flt);
-            add_memref_descriptor(refs, inp);
-            add_memref_descriptor(refs, out);
-            refs.push_back(out);
-
-            mod->replace_instruction(ins, conv, refs);
-        }
-    }
-
-    void apply()
-    {
-        init();
-        for(auto it : iterator_for(*mod))
-        {
-            if(it->name() == "convolution")
-            {
-                replace_conv_op(it);
-            }
-        }
-    }
-};
-
-void mlir_conv::apply(module& m) const { mlir_apply{&m, this}.apply(); }
-
-} // namespace gpu
-} // namespace MIGRAPHX_INLINE_NS
-} // namespace migraphx

src/targets/gpu/target.cpp

@@ -53,10 +53,10 @@
 #include <migraphx/gpu/compile_ops.hpp>
 #include <migraphx/gpu/concat_gpu_opt.hpp>
 #include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/fuse_mlir.hpp>
 #include <migraphx/gpu/fuse_ops.hpp>
 #include <migraphx/gpu/prefuse_ops.hpp>
 #include <migraphx/gpu/lowering.hpp>
-#include <migraphx/gpu/mlir_conv.hpp>
 #include <migraphx/gpu/pack_int8_args.hpp>
 #include <migraphx/gpu/schedule_model.hpp>
 #include <migraphx/gpu/sync_device.hpp>
@@ -128,7 +128,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         dead_code_elimination{},
         enable_pass(not enabled(MIGRAPHX_DISABLE_POINTWISE_FUSION{}), fuse_pointwise{}),
         dead_code_elimination{},
-        mlir_conv{&ctx},
+        fuse_mlir{&ctx},
+        dead_code_elimination{},
         lowering{&ctx, options.offload_copy},
         eliminate_contiguous{"gpu::contiguous"},
         dead_code_elimination{},

test/gpu/mlir.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/gpu/mlir.hpp>
+#include <migraphx/gpu/target.hpp>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/write_literals.hpp>
+#include <migraphx/ref/target.hpp>
+#include <migraphx/module.hpp>
+#include <migraphx/program.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/stringutils.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/verify_args.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/functional.hpp>
+#include <test.hpp>
+
+using migraphx::trim;
+
+// m test_gpu_mlir && ./bin/test_gpu_mlir
+
+struct mlir_gpu_target : migraphx::gpu::target
+{
+    std::string name() const { return "mlir"; }
+    std::vector<migraphx::pass> get_passes(migraphx::context& gctx,
+                                           const migraphx::compile_options&) const
+    {
+        auto& ctx = migraphx::any_cast<migraphx::gpu::context>(gctx);
+        return {migraphx::gpu::write_literals{&ctx}};
+    }
+};
+
+std::string encode(const std::string& s)
+{
+    std::stringstream ss;
+    bool prespace = false;
+    for(auto c : s)
+    {
+        if(std::isspace(c) != 0)
+        {
+            if(not prespace)
+                ss << "  ";
+            prespace = true;
+        }
+        else if(std::isprint(c) != 0)
+        {
+            ss << c;
+            prespace = false;
+        }
+    }
+    return migraphx::trim(ss.str());
+}
+
+migraphx::program create_program_from_mlir(const migraphx::module& mmlir)
+{
+    migraphx::program p;
+    auto* mm   = p.get_main_module();
+    auto names = mmlir.get_parameter_names();
+    std::vector<migraphx::instruction_ref> inputs;
+    std::transform(names.begin(), names.end(), std::back_inserter(inputs), [&](const auto& name) {
+        return mm->add_parameter(name, mmlir.get_parameter_shape(name));
+    });
+    std::sort(inputs.begin(), inputs.end(), migraphx::by(std::less<>{}, [](auto ins) {
+                  return to_string(ins->get_operator());
+              }));
+    inputs.push_back(mm->add_parameter("output", mmlir.get_output_shapes().front()));
+
+    migraphx::gpu::context ctx;
+    migraphx::gpu::insert_mlir(*mm, mm->end(), compile_mlir(ctx, mmlir), inputs);
+    return p;
+}
+
+migraphx::parameter_map generate_params(const migraphx::program& p)
+{
+    migraphx::parameter_map m;
+    std::size_t i = 0;
+    for(auto&& x : p.get_parameter_shapes())
+    {
+        // m[x.first] = migraphx::fill_argument(x.second, 1);
+        m[x.first] = migraphx::generate_argument(x.second, i++);
+    }
+    return m;
+}
+
+migraphx::argument run_gpu(migraphx::program p, const migraphx::parameter_map& inputs)
+{
+    mlir_gpu_target t;
+    p.compile(t);
+    migraphx::parameter_map m;
+    for(auto&& input : inputs)
+    {
+        m[input.first] = t.copy_to(input.second);
+    }
+    for(auto&& x : p.get_parameter_shapes())
+    {
+        if(m.count(x.first) == 0)
+        {
+            m[x.first] = t.allocate(x.second);
+        }
+    }
+    return t.copy_from(p.eval(m).front());
+}
+
+migraphx::argument run_ref(migraphx::program p, const migraphx::parameter_map& inputs)
+{
+    p.compile(migraphx::ref::target{});
+    return p.eval(inputs).front();
+}
+
+bool verify_mlir(const migraphx::module& mmlir)
+{
+    migraphx::program ref;
+    ref.get_main_module()->insert_instructions(ref.get_main_module()->end(), &mmlir);
+
+    auto inputs = generate_params(ref);
+
+    auto mlir = create_program_from_mlir(mmlir);
+    return migraphx::verify_args("mlir", run_ref(ref, inputs), run_gpu(mlir, inputs));
+}
+
+TEST_CASE(conv)
+{
+    const std::string mlir_output = R"__migraphx__(
+module {
+  func @main(%arg0: tensor<2x8x3x3xf32>, %arg1: tensor<1x8x4x4xf32>) -> tensor<1x2x2x2xf32> attributes {kernel = "mixr"} {
+    %0 = migraphx.convolution(%arg1, %arg0) {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1]} : (tensor<1x8x4x4xf32>, tensor<2x8x3x3xf32>) -> tensor<1x2x2x2xf32>
+    return %0 : tensor<1x2x2x2xf32>
+  }
+}
+)__migraphx__";
+    migraphx::module m;
+    auto x    = m.add_parameter("x", {migraphx::shape::float_type, {1, 8, 4, 4}});
+    auto w    = m.add_parameter("w", {migraphx::shape::float_type, {2, 8, 3, 3}});
+    auto conv = m.add_instruction(migraphx::make_op("convolution"), x, w);
+    m.add_return({conv});
+    auto s = migraphx::gpu::dump_mlir(m);
+    // Skip test if MLIR is not enabled
+    if(s.empty())
+        return;
+    CHECK(encode(s) == encode(mlir_output));
+    EXPECT(verify_mlir(m));
+}
+
+TEST_CASE(conv_add_relu)
+{
+    const std::string mlir_output = R"__migraphx__(
+module {
+  func @main(%arg0: tensor<1x2x2x2xf32>, %arg1: tensor<2x8x3x3xf32>, %arg2: tensor<1x8x4x4xf32>) -> tensor<1x2x2x2xf32> attributes {kernel = "mixr"} {
+    %0 = migraphx.convolution(%arg2, %arg1) {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1]} : (tensor<1x8x4x4xf32>, tensor<2x8x3x3xf32>) -> tensor<1x2x2x2xf32>
+    %1 = migraphx.add(%0, %arg0) : (tensor<1x2x2x2xf32>, tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
+    %2 = migraphx.relu(%1) : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
+    return %2 : tensor<1x2x2x2xf32>
+  }
+}
+)__migraphx__";
+    migraphx::module m;
+    auto x    = m.add_parameter("x", {migraphx::shape::float_type, {1, 8, 4, 4}});
+    auto w    = m.add_parameter("w", {migraphx::shape::float_type, {2, 8, 3, 3}});
+    auto b    = m.add_parameter("b", {migraphx::shape::float_type, {1, 2, 2, 2}});
+    auto conv = m.add_instruction(migraphx::make_op("convolution"), x, w);
+    auto add  = m.add_instruction(migraphx::make_op("add"), conv, b);
+    auto relu = m.add_instruction(migraphx::make_op("relu"), add);
+    m.add_return({relu});
+    auto s = migraphx::gpu::dump_mlir(m);
+    // Skip test if MLIR is not enabled
+    if(s.empty())
+        return;
+    CHECK(encode(s) == encode(mlir_output));
+    EXPECT(verify_mlir(m));
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }